An innovative approach in fault diagnostics of steam turbine performance using ANN and M-SVM algorithms

Abstract

In this study, an artificial neural network (ANN) and multi-support vector machine (M-SVM) models are developed to predict seven kinds of faults associated with turbine of thermal power plant. Information about changes with respect to base (design / performance guaranteed) values in seven key parameters/indices namely throttle flow, throttle pressure, first stage pressure, hot reheat pressure, cross-over pressure, HP efficiency and IP efficiency enable proper interpretation of different faults such as solid particle erosion (SPE ) in HP turbine, IP turbine, deposits and peening in HP turbine, deposits and peening in IP turbine, deposits in IP turbine, rubs in IP turbine and deposits and damages in IP turbine. All the fault types are diagnosed based on observation of seven process parameters/ indices over a long time. The trends of changes - either increasing or decreasing direction – with respect to base values are fed as inputs to the networks and their output identifies the type of fault. A comparison between the classifier models is discussed with the proposed action plan to improve the smooth running of the machines